Abstract: Many chemical processes have the characteristics of strong nonlinearity, complex mechanism and multiple operating conditions. Aiming at the problem that traditional soft sensor model can’t fully describe the process characteristics, which leads to the low prediction accuracy, a multi-model fusion soft sensor modeling method based on FCM-ABC-MKRVM is proposed. Firstly, the fuzzy C-means (FCM) clustering algorithm was used to divide the training samples into several subclasses and the clustering centers of each subclass were determined. Then, the multi-kernel relevance vector machine (MKRVM) sub-models were established by training each subclass samples. The kernel parameter and weight factor were optimized by artificial bee colony (ABC) algorithm. In the stage of the model prediction, the membership values between the test samples and the cluster centers were calculated as the weight coefficients of the output values of sub models. The final prediction output was obtained by the multi-model fusion. The proposed modeling method was applied to develop polypropylene melt index soft sensor. The result shows that the melt index soft sensor model based on FCM-ABC-MKRVM has better predicting accuracy compared with the MKRVM model and the ABC-MKRVM model. The proposed modeling method could provide guidance for online predicting of the quality index of chemical process under complex multi-operating modes.

Key words: Fuzzy C-means clustering, artificial bee colony algorithm, multi-kernel relevance vector machine, multi-model fusion, soft sensor, melt index

摘要： 许多化工过程具有强非线性、机理复杂和多工况等特点，针对传统软测量模型无法全面描述过程特性而导致模型预测精度较低的问题，提出一种FCM-ABC-MKRVM多模型融合软测量建模方法。首先采用模糊C均值聚类算法（FCM）将训练样本划分为多个子类，并确定各子类的聚类中心；然后通过训练各子类样本建立多核相关向量机（MKRVM）子模型，其中采用人工蜂群算法（ABC）优化核函数参数和组合权重因子；在模型预测阶段，计算测试样本与各聚类中心的隶属度值，并作为各子模型输出值的加权系数，通过多模型融合得到最终的模型预测输出。将该建模方法应用于聚丙烯熔融指数软测量研究中，仿真结果表明：与MKRVM模型和ABC-MKRVM模型相比，基于FCM-ABC-MKRVM多模型融合的熔融指数软测量模型具有更佳的预测精度，可以为复杂多工况化工过程的产品质量指标在线预测提供指导作用。

关键词: 模糊C均值聚类, 人工蜂群算法, 多核相关向量机, 多模型融合, 软测量, 熔融指数